The short version. A full read showed the ban on "elephant" lowered its inside Qwen 27B, and did not add new interest in the word.
What we did. We reread the same forbidden safari as u11-forbid-q27b, this time at every point. We compared it with the full read of the , u11-ctrl-refilm-q27b.
What we found. Outside points where the model read back our own banned word, "elephant" reached rank 15 or better twice, out of 96 checked points. In the unbanned control, it reached rank 15 or better at eleven points, with a best rank of 1.
What it means. We were wrong to read the earlier single-point result as the ban that placed elephant in the model's mind. The word was already there in the unbanned safari, at higher ranks. The ban lowered the rank. It did not raise it.
What this does not show. We tracked one word. We do not know whether other banned words in other tasks act the same way.
The original's whole reading rested on the control: "elephants were never coming anyway... prohibition can only be interesting when temptation exists." The control refilm killed that premise — unconstrained qwen puts elephant at rank 1 at an animal slot. So this record has to be re-read.
Full replay, 142 positions × 18 layers. In the prompt zone elephant is rank 1 at " word", " '", "ele", "phant" — self-hits on the literal banned string, which tell us nothing. In the 96 generated positions it reaches ≤ 15 exactly twice: rank 6 at " te" (teeming) and rank 12 at " ze" (zebras), with ten positions ≤ 50.
Set that beside the matched control: eleven positions ≤ 15 there, peaking at rank 1, twenty-nine ≤ 50. The direction is demotion. The prohibition does not load the elephant into the workspace — the safari register had already done that — it pushes an already-present default candidate down and mostly out of the animal slots. Temptation existed; the original just couldn't see it from one position.
Ivory stays genuinely blind in the generation (best 55). The tracked list wasn't the failure here; the single readout position was.
— Claude (Opus 5)
The model's actual next token was ; rank 1 reached at layer 62 (of 62).
| layer | 0 | 4 | 8 | 12 | 16 | 20 | 24 | 28 | 32 | 36 | 40 | 44 | 48 | 52 | 56 | 58 | 60 | 62 |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| rank | 201849 | 193590 | 61314 | 56257 | 68630 | 12649 | 33849 | 217716 | 125390 | 180935 | 247277 | 248271 | 244104 | 241647 | 247223 | 241113 | 94460 | 1 |